fix: join autotune timer thread when training raises

python/fasttext_module/fasttext/tests/test_autotune_errors.py

@@ -0,0 +1,26 @@
+# SPDX-FileContributor: Arthit Suriyawongkul
+# SPDX-FileCopyrightText: 2026-present, fasttext-community
+# SPDX-FileType: SOURCE
+# SPDX-License-Identifier: MIT
+
+"""Autotune errors must raise, not terminate the process."""
+
+import pytest
+
+from .helpers import build_supervised_model, get_random_data
+
+
+def test_autotune_error_raises(tmp_path):
+    data = get_random_data(3000, max_vocab_size=600)
+    valid = tmp_path / "valid.txt"
+    valid.write_text("".join(f"__label__{line}\n" for line in data))
+    kwargs = {
+        # thread=12: thread <= 10 leaves the input matrix partly uninitialized.
+        "thread": 12,
+        "verbose": 0,
+        "autotuneValidationFile": str(valid),
+        "autotuneMetric": "f1:__label__missing",  # fails after the first trial
+        "autotuneDuration": 60,  # long enough for slow runners to reach it
+    }
+    with pytest.raises(RuntimeError, match="Unknown autotune metric label"):
+        build_supervised_model(data, kwargs)

src/autotune.cc

@@ -215,6 +215,15 @@ Autotune::Autotune(const std::shared_ptr<FastText>& fastText)
       strategy_(),
       timer_() {}
 
+Autotune::~Autotune() noexcept {
+  // An exception leaving train() skips its timer join, and destroying a
+  // joinable std::thread calls std::terminate().
+  if (timer_.joinable()) {
+    continueTraining_ = false;
+    timer_.join();
+  }
+}
+
 void Autotune::printInfo(double maxDuration) {
   double progress = elapsed_ * 100 / maxDuration;
   progress = std::min(progress, 100.0);

src/autotune.h

@@ -81,7 +81,7 @@ class Autotune {
   Autotune(Autotune&&) = delete;
   Autotune& operator=(const Autotune&) = delete;
   Autotune& operator=(Autotune&&) = delete;
-  ~Autotune() noexcept = default;
+  ~Autotune() noexcept;
 
   void train(const Args& args);
 };